SPHERE (.sph) to WAV Converter

About this tool

NIST SPHERE is the container almost every classic speech corpus ships in — TIMIT, Switchboard, Fisher, WSJ, CALLHOME, TEDLIUM. A .sph file is a fixed-size block of ASCII header text followed by raw interleaved samples, and nothing on a modern desktop plays it: the extension is often .wav even though the bytes are not RIFF, the samples are frequently big-endian, and telephone corpora store 8-bit mu-law or A-law instead of linear PCM.

This converter reads the header field by field, applies what it says, and writes a standard RIFF/WAVE file you can open anywhere. Paste the .sph bytes as base64 or hex (or a data: URI), pick what you want back, and the conversion runs locally in your browser through WebAssembly — the recording never leaves the machine.

What it does with the header:

• sample_byte_format — 10 means big-endian, so every 16-bit sample is byte-swapped into the little-endian order WAV requires. 01 passes through untouched. Force it with Byte order when a corpus ships a wrong value and the result sounds like noise.
• sample_coding — pcm is copied, ulaw / mu-law and alaw are expanded to 16-bit PCM by G.711 decoding so ordinary players handle them.
• channel_count — two-channel conversation recordings can be split (Channel 1 only / Channel 2 only) or averaged into a mono downmix.
• sample_count, sample_n_bytes — used to locate and bound the sample data; a file whose declared sample count exceeds the bytes actually present is reported as truncated rather than quietly half-converted.

Worked example

The demo file behind the “Convert to WAV” chip has this header, padded to 256 bytes and followed by 40 bytes of samples:

NIST_1A
    256
sample_rate -i 8000
channel_count -i 1
sample_n_bytes -i 2
sample_byte_format -s2 10
sample_coding -s3 pcm
sample_count -i 20
end_head

Its first sample pair on disk is 00 00 0f e0 — big-endian. With the defaults (16-bit PCM, all channels, WAV container) the output is an 84-byte file: a 44-byte RIFF header plus the 40 byte-swapped sample bytes, returned as

data:audio/wav;base64,UklGRkwAAABXQVZFZm10IBAAAAABAAEAQB8AAIA+AAACABAAZGF0YSgAAAAA…

Switch Container to Raw, headerless samples and Return to Hex audio bytes and you get the samples on their own, now little-endian:

0000e00fe11d58280b2e4c2e1329ff1e4111780184f146e36dd841d27fd13cd6eadf63ed0ffd130d

Switch Return to Header report and you get the parsed field table instead of audio: every header line with its type token, then the derived sample rate, channel count, coding, byte order, frame count and duration, plus what the conversion would produce (container, encoding, size). For Raw, headerless samples the report ends with a ready-to-run re-import command such as ffmpeg -f s16le -ar 8000 -ac 1 -i out.raw out.wav, because headerless audio carries none of those parameters with it.

Limits and edge cases

• Decoded input is capped at 6 MiB and produced audio at 12 MiB; the hex rendering is capped at 4 MiB of audio because it doubles again as text. Use Start sample frame / Max sample frames to excerpt a long recording — one frame is one sample per channel, so seconds × sample_rate gives the frame index.
• Shorten-compressed payloads (sample_coding: pcm,embedded-shorten-v1.09 and friends) are detected and reported, not decoded. Decompress those with a desktop converter that bundles a shorten decoder, then convert the uncompressed file here.
• The sample rate is never changed — this tool re-containers audio, it does not resample.
• 8-bit output is written unsigned, because that is what WAV requires of 8-bit PCM; 16-, 24- and 32-bit output stays signed little-endian.
• A header with no sample_byte_format and samples wider than one byte is an error rather than a guess; set Byte order explicitly.
• Files with a header size other than the usual 1024 bytes are fine — the size on line 2 is what is honoured.

FAQ